Technology Awards

FRVT 1:1 Verification

These comments provided by Neurotechnology are based on NIST FRVT 1:1 Verification ongoing evaluation results published on February 23, 2022.

Neurotechnology algorithm has been ranked in the top 3% of most accurate results for border control supervised (Visa Border, Border) and unsupervised (Kiosk) scenarios among 702 submissions by 255 providers. The submission was also among the top 3% of algorithms for accuracy with face masks from a total of 319 entries.

NIST has started the FRVT Ongoing evaluation of face verification algorithms in February 2017. Since then, 728 algorithms from 267 vendors were tested on 7 datasets of different complexity. Each vendor can submit multiple algorithms once every three months, but the report takes into account only 2 latest submissions per vendor. Thus, only 396 algorithms are analyzed in the current results table from NIST.

The chart below shows the accuracy of all tested algorithms, with red line showing Neurotechnology algorithm.

Neurotechnology algorithm performance in FRVT 1:1

The horizontal axis on the chart corresponds to the performed tests, which are described below:

  1. FRVT 1:1 VISA Photos
  2. FRVT 1:1 MUGSHOT Photos
  3. FRVT 1:1 MUGSHOT Photos ΔT >12 YRS
  4. FRVT 1:1 VISABORDER Photos
  5. FRVT 1:1 BORDER Photos
  6. FRVT 1:1 WILD Photos
  7. FRVT 1:1 KIOSK Photos
  8. FRVT Face Mask Effects VISABORDER Photos NOT MASKED
  9. FRVT Face Mask Effects VISABORDER Photos MASKED PROBE

Visa

The dataset represents typical photos of US visa applicants. According to the dataset description, the face images are generally high quality. Part of the images are live capture and the other part is photographed from paper photos.

Neurotechnology algorithm accuracy in this scenario was 0.45% FNMR at 0.0001% FMR. The most accurate contender showed 0.13% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • All face templates were compared between each other.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).
  • The comparisons were fully zero-effort, meaning there was no pre-grouping by gender, age or other covariates.

VISA dataset summary

  • More than 100,000 unique persons.
  • Part of the images are live capture and part are digitized paper photos.
  • High quality images, with conformance with the ISO/IEC 19794-5 Full Frontal image type. Subjects' pose is generally excellent.
  • Image size: 252x300 pixels.
  • Interocular distance (IOD): 69 pixels.
  • Subjects from more than 100 countries are represented, with significant imbalance due to visa issuance patterns.
  • All ages are represented, including children, with imbalance due to visa issuance demand.

Mugshot

The dataset represents typical photos of suspects taken by law enforcement officers. The photos were taken in a controlled environment thus their quality is high.

Neurotechnology algorithm accuracy in this scenario was 0.26% FNMR at 0.001% FMR. The most accurate contender showed 0.21% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • All face templates were compared between each other.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).
  • Gender attribute was used to exclude comparisons between different gender subjects.

MUGSHOT dataset summary

  • More than 1,000,000 unique persons.
  • All images were live capture.
  • High quality images, with conformance with the ISO/IEC 19794-5 Full Frontal image type. Subjects' pose is generally excellent.
  • Variable image sizes.
  • Interocular distance (IOD): from 34 to 297 pixels, with 113 pixels mean and 105 pixels median values.
  • All subjects are from the United States.
  • All subjects are adults.

Mugshot ΔT ≥12 years

The test was performed using the same dataset as in the MUGSHOT test above. The comparisons were performed only between mugshots representing same or different persons with age difference of 12 or more years between them.

Neurotechnology algorithm accuracy in this scenario was 0.36% FNMR at 0.001% FMR. The most accurate contender showed 0.20% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • Age attribute was used to exclude template pairs, which represented same or different persons with age difference of less than 12 years at the capture moment, from the comparison.
  • Gender attribute was used to exclude comparisons between different gender subjects.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).

Visa Border

The test was performed by comparing high-quality images from the VISA dataset against lower quality images from the BORDER dataset. Both datasets include subjects from more than 100 countries, with specific imbalances due to visa issuance patterns and border-crossing demographics correspondingly.

Neurotechnology algorithm accuracy in this scenario was 0.37% FNMR at 0.0001% FMR. The most accurate contender showed 0.21% FNMR at the same FMR.

Scenario overview

  • One image from Visa or Border datasets was used to create one face template.
  • Gender attribute was used to exclude comparisons between different gender subjects.
  • Each pre-filtered template from the Visa dataset was compared against each pre-filtered template from the Border dataset
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).

Border

The dataset is made of images obtained at border crossing. Images are taken by operators with webcams oriented towards cooperating subjects. The quality of the images may vary due to time-constrained capture and different illumination conditions.

Neurotechnology algorithm accuracy in this scenario was 0.65% FNMR at 0.0001% FMR. The most accurate contender showed 0.40% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • Gender attribute was used to exclude comparisons between different gender subjects.
  • All pre-filtered face templates were compared between each other.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).

Border dataset summary

  • More than 1,000,000 unique persons.
  • All images were live capture
  • Variable image quality. There are role, pitch and yaw angle variations. There is some perspective distortion due to close range images. Background illumination is sometimes strong, so faces may be under-exposed. Some faces are partially cropped.
  • Interocular distance (IOD): 38 pixels mean.
  • Subjects from more than 100 countries are represented, with significant imbalance due to immigration patterns.
  • All subjects are adults, with age distribution imbalance due to immigration patterns.

Wild

The dataset includes cropped photojournalism-style images of subjects, who do not make any efforts to be captured, thus their poses may vary very widely, and their faces can be occluded, including hair and hands. Image quality and resolution can also vary as the images are captured in random environment

Neurotechnology algorithm accuracy in this scenario was 3.01% FNMR at 0.001% FMR. The most accurate contender showed 2.93% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • All face templates were compared between each other.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).
  • The comparisons were fully zero-effort, meaning there was no pre-grouping by gender, age or other covariates.

WILD dataset summary

  • More than 10,000 unique persons.
  • All images were live capture.
  • Variable image quality. There is wide yaw and pitch pose variation. Faces can be occluded, including hair and hands.
  • Variable image sizes. Images were post-processed using a variable but generally tight crop of the head.
  • All subjects are adults.

Kiosk

The dataset is made from images. which were captured from cooperating subjects in a partially controlled environment. Images' quality varied due to an automated, non-supervised capture method.

Neurotechnology algorithm accuracy in this scenario was 6.28% FNMR at 0.001% FMR. The most accurate contender showed 4.61% FNMR at the same FMR.

Scenario overview

  • One image from the dataset was used to create one face template.
  • All face templates were compared between each other.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).
  • The comparisons were fully zero-effort, meaning there was no pre-grouping by gender, age or other covariates.

KIOSK dataset summary

  • All images were live capture.
  • Many images have a considerable downward pitch angle.
  • In some images, faces are partially cropped.
  • Some images have other faces in the background.

Visa Border Not masked and Masked Probe

These tests are part of separate FRVT Face Mask Effects evaluation, which used VISA and BORDER datasets from the FRVT 1:1 evaluation.

VISA BORDER NOT MASKED scenario

  • The testing scenario is the same as the VISA BORDER from the FRVT 1:1, except that the resulting FNMR value is obtained using a different FMR value.
  • Neurotechnology algorithm accuracy in this scenario was 0.28% FNMR at 0.0001% FMR.
  • The most accurate contender showed 0.18% FNMR at the same FMR.

VISA BORDER MASKED PROBE scenario

  • One original image from the VISA dataset was used to create one face template.
  • Each original image from the BORDER dataset was post-processed in five different ways to create synthetic lower face part occlusions as if the subjects wear face masks (see the figure below). One face template was created from each variant of the modified image.
  • Gender attribute was used to exclude comparisons between different gender subjects.
  • Each pre-filtered template from the VISA dataset was compared against each pre-filtered template from the BORDER dataset with synthetic face masks.
  • Each comparison score was obtained by comparing two different templates (simple one-to-one verification).
  • Neurotechnology algorithm accuracy in this scenario was 1.80% FNMR at 0.0001% FMR.
  • The most accurate contender showed 0.85% FNMR at the same FMR.
Sample synthetic masks
Examples of synthetically-generated face masks
Image credit: B. Hayes, M. Ngan/NIST.
Examples of synthetically-generated face masks
Image credit: B. Hayes, M. Ngan/NIST.
Sample synthetic masks
Click to zoom
Representatives
Neurotechnology Distributors Map Ex-Cle S.A - representative in Argentina FingerSec do Brasil - distributor in Brazil (web site in Portuguese) Distributors in Chile Neurotechnology's Chinese Office (web site in Chinese) Security Systems Ltda - distributor in Colombia (web site in Spanish) General Security El Salvador - distributor in El Salvador (web site in Spanish) Infokey Software Solutions - distributor in Greece (web site in Greek and English) India Branch - Neurotechnology Lab India Fulcrum Biometrics India Pvt. Ltd. - distributor in India Biometric srl - distributor in Italy (web site in Italian) Software Sources Ltd - distributor in Israel Bruce and Brian Co., LTD. - distributor in Korea (web site in Korean) Biosec Solutions - distributor in Nigeria Digital Data Systems (DDS Biometrics) - distributor in Pakistan Ex-Cle S.A - distributor in Paraguay Digital Works - distributor in Peru DigiFace Solutions - distributor in Singapore Fingerprint i.t. - distributor in South Africa Sri Lanka Branch - Neurotechnology Lab Delaney Biometrics - distributor in the UK Fulcrum Biometrics - representative in the USA
Follow us
Facebook icon   LinkedIn icon   Twitter icon   Youtube icon   Email newsletter icon
Copyright © 1998 - 2022 Neurotechnology | Terms & Conditions | Privacy Policy | Career